A framework for ancient and machine-printed manuscripts categorization

Document image understanding (DIU) has attracted a lot of attention and became an of active fields of research. Although, the ultimate goal of DIU is extracting textual information of a document image, many steps are involved in a such a process such as categorization, segmentation and layout analysis. All of these steps are needed in order to obtain an accurate result from character recognition or word recognition of a document image. One of the important steps in DIU is document image categorization (DIC) that is needed in many situations such as document image written or printed in more than one script, font or language. This step provides useful information for recognition system and helps in reducing its error by allowing to incorporate a category-specific Optical Character Recognition (OCR) system or word recognition (WR) system. This research focuses on the problem of DIC in different categories of scripts, styles and languages and establishes a framework for flexible representation and feature extraction that can be adapted to many DIC problem. The current methods for DIC have many limitations and drawbacks that restrict the practical usage of these methods. We proposed an efficient framework for categorization of document image based on patch representation and Non-negative Matrix Factorization (NMF). This framework is flexible and can be adapted to different categorization problem. Many methods exist for script identification of document image but few of them addressed the problem in handwritten manuscripts and they have many limitations and drawbacks. Therefore, our first goal is to introduce a novel method for script identification of ancient manuscripts. The proposed method is based on patch representation in which the patches are extracted using skeleton map of a document images. This representation overcomes the limitation of the current methods about the fixed level of layout. The proposed feature extraction scheme based on Projective Non-negative Matrix Factorization (PNMF) is robust against noise and handwriting variation and can be used for different scripts. The proposed method has higher performance compared to state of the art methods and can be applied to different levels of layout. The current methods for font (style) identification are mostly proposed to be applied on machine-printed document image and many of them can only be used for a specific level of layout. Therefore, we proposed new method for font and style identification of printed and handwritten manuscripts based on patch representation and Non-negative Matrix Tri-Factorization (NMTF). The images are represented by overlapping patches obtained from the foreground pixels. The position of these patches are set based on skeleton map to reduce the number of patches. Non-Negative Matrix Tri-Factorization is used to learn bases from each fonts (style) and then these bases are used to classify a new image based on minimum representation error. The proposed method can easily be extended to new fonts as the bases for each font are learned separately from the other fonts. This method is tested on two datasets of machine-printed and ancient manuscript and the results confirmed its performance compared to the state of the art methods. Finally, we proposed a novel method for language identification of printed and handwritten manuscripts based on patch representation and Non-negative Matrix Tri-Factorization (NMTF). The current methods for language identification are based on textual data obtained by OCR engine or images data through coding and comparing with textual data. The OCR based method needs lots of processing and the current image based method are not applicable to cursive scripts such as Arabic. In this work we introduced a new method for language identification of machine-printed and handwritten manuscripts based on patch representation and NMTF. The patch representation provides the component of the Arabic script (letters) that can not be extracted simply by segmentation methods. Then NMTF is used for dictionary learning and generating codebooks that will be used to represent document image with a histogram. The proposed method is tested on two datasets of machine-printed and handwritten manuscripts and compared to n-gram features (text-based), texture features and codebook features (imagebased) to validate the performance. The above proposed methods are robust against variation in handwritings, changes in the font (handwriting style) and presence of degradation and are flexible that can be used to various levels of layout (from a textline to paragraph). The methods in this research have been tested on datasets of handwritten and machine-printed manuscripts and compared to state-of-the-art methods. All of the evaluations show the efficiency, robustness and flexibility of the proposed methods for categorization of document image. As mentioned before the proposed strategies provide a framework for efficient and flexible representation and feature extraction for document image categorization. This frame work can be applied to different levels of layout, the information from different levels of layout can be merged and mixed and this framework can be extended to more complex situations and different tasks